fs/btrfs/ctree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/sched.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "print-tree.h"
  24
  25 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
  26                       *root, struct btrfs_path *path, int level);
  27 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
  28                       *root, struct btrfs_key *ins_key,
  29                       struct btrfs_path *path, int data_size, int extend);
  30 static int push_node_left(struct btrfs_trans_handle *trans,
  31                           struct btrfs_root *root, struct extent_buffer *dst,
  32                           struct extent_buffer *src);
  33 static int balance_node_right(struct btrfs_trans_handle *trans,
  34                               struct btrfs_root *root,
  35                               struct extent_buffer *dst_buf,
  36                               struct extent_buffer *src_buf);
  37 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
  38                    struct btrfs_path *path, int level, int slot);
  39
  40 inline void btrfs_init_path(struct btrfs_path *p)
  41 {
  42         memset(p, 0, sizeof(*p));
  43 }
  44
  45 struct btrfs_path *btrfs_alloc_path(void)
  46 {
  47         struct btrfs_path *path;
  48         path = kmem_cache_alloc(btrfs_path_cachep, GFP_NOFS);
  49         if (path) {
  50                 btrfs_init_path(path);
  51                 path->reada = 1;
  52         }
  53         return path;
  54 }
  55
  56 void btrfs_free_path(struct btrfs_path *p)
  57 {
  58         btrfs_release_path(NULL, p);
  59         kmem_cache_free(btrfs_path_cachep, p);
  60 }
  61
  62 void btrfs_release_path(struct btrfs_root *root, struct btrfs_path *p)
  63 {
  64         int i;
  65         for (i = 0; i < BTRFS_MAX_LEVEL; i++) {
  66                 if (!p->nodes[i])
  67                         break;
  68                 free_extent_buffer(p->nodes[i]);
  69         }
  70         memset(p, 0, sizeof(*p));
  71 }
  72
  73 static int __btrfs_cow_block(struct btrfs_trans_handle *trans,
  74                              struct btrfs_root *root,
  75                              struct extent_buffer *buf,
  76                              struct extent_buffer *parent, int parent_slot,
  77                              struct extent_buffer **cow_ret,
  78                              u64 search_start, u64 empty_size)
  79 {
  80         u64 root_gen;
  81         struct extent_buffer *cow;
  82         u32 nritems;
  83         int ret = 0;
  84         int different_trans = 0;
  85         int level;
  86         struct btrfs_key first_key;
  87
  88         if (root->ref_cows) {
  89                 root_gen = trans->transid;
  90         } else {
  91                 root_gen = 0;
  92         }
  93
  94         WARN_ON(root->ref_cows && trans->transid !=
  95                 root->fs_info->running_transaction->transid);
  96         WARN_ON(root->ref_cows && trans->transid != root->last_trans);
  97
  98         level = btrfs_header_level(buf);
  99         nritems = btrfs_header_nritems(buf);
 100         if (nritems) {
 101                 if (level == 0)
 102                         btrfs_item_key_to_cpu(buf, &first_key, 0);
 103                 else
 104                         btrfs_node_key_to_cpu(buf, &first_key, 0);
 105         } else {
 106                 first_key.objectid = 0;
 107         }
 108         cow = __btrfs_alloc_free_block(trans, root, buf->len,
 109                                      root->root_key.objectid,
 110                                      root_gen, first_key.objectid, level,
 111                                      search_start, empty_size);
 112         if (IS_ERR(cow))
 113                 return PTR_ERR(cow);
 114
 115         copy_extent_buffer(cow, buf, 0, 0, cow->len);
 116         btrfs_set_header_bytenr(cow, cow->start);
 117         btrfs_set_header_generation(cow, trans->transid);
 118         btrfs_set_header_owner(cow, root->root_key.objectid);
 119
 120         WARN_ON(btrfs_header_generation(buf) > trans->transid);
 121         if (btrfs_header_generation(buf) != trans->transid) {
 122                 different_trans = 1;
 123                 ret = btrfs_inc_ref(trans, root, buf);
 124                 if (ret)
 125                         return ret;
 126         } else {
 127                 clean_tree_block(trans, root, buf);
 128         }
 129
 130         if (buf == root->node) {
 131                 root_gen = btrfs_header_generation(buf);
 132                 root->node = cow;
 133                 extent_buffer_get(cow);
 134                 if (buf != root->commit_root) {
 135                         btrfs_free_extent(trans, root, buf->start,
 136                                           buf->len, root->root_key.objectid,
 137                                           root_gen, 0, 0, 1);
 138                 }
 139                 free_extent_buffer(buf);
 140         } else {
 141                 root_gen = btrfs_header_generation(parent);
 142                 btrfs_set_node_blockptr(parent, parent_slot,
 143                                         cow->start);
 144                 WARN_ON(trans->transid == 0);
 145                 btrfs_set_node_ptr_generation(parent, parent_slot,
 146                                               trans->transid);
 147                 btrfs_mark_buffer_dirty(parent);
 148                 WARN_ON(btrfs_header_generation(parent) != trans->transid);
 149                 btrfs_free_extent(trans, root, buf->start, buf->len,
 150                                   btrfs_header_owner(parent), root_gen,
 151                                   0, 0, 1);
 152         }
 153         free_extent_buffer(buf);
 154         btrfs_mark_buffer_dirty(cow);
 155         *cow_ret = cow;
 156         return 0;
 157 }
 158
 159 int btrfs_cow_block(struct btrfs_trans_handle *trans,
 160                     struct btrfs_root *root, struct extent_buffer *buf,
 161                     struct extent_buffer *parent, int parent_slot,
 162                     struct extent_buffer **cow_ret)
 163 {
 164         u64 search_start;
 165         int ret;
 166         if (trans->transaction != root->fs_info->running_transaction) {
 167                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 168                        root->fs_info->running_transaction->transid);
 169                 WARN_ON(1);
 170         }
 171         if (trans->transid != root->fs_info->generation) {
 172                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 173                        root->fs_info->generation);
 174                 WARN_ON(1);
 175         }
 176         if (btrfs_header_generation(buf) == trans->transid) {
 177                 *cow_ret = buf;
 178                 return 0;
 179         }
 180
 181         search_start = buf->start & ~((u64)BTRFS_BLOCK_GROUP_SIZE - 1);
 182         ret = __btrfs_cow_block(trans, root, buf, parent,
 183                                  parent_slot, cow_ret, search_start, 0);
 184         return ret;
 185 }
 186
 187 static int close_blocks(u64 blocknr, u64 other, u32 blocksize)
 188 {
 189         if (blocknr < other && other - (blocknr + blocksize) < 32768)
 190                 return 1;
 191         if (blocknr > other && blocknr - (other + blocksize) < 32768)
 192                 return 1;
 193         return 0;
 194 }
 195
 196 /*
 197  * compare two keys in a memcmp fashion
 198  */
 199 static int comp_keys(struct btrfs_disk_key *disk, struct btrfs_key *k2)
 200 {
 201         struct btrfs_key k1;
 202
 203         btrfs_disk_key_to_cpu(&k1, disk);
 204
 205         if (k1.objectid > k2->objectid)
 206                 return 1;
 207         if (k1.objectid < k2->objectid)
 208                 return -1;
 209         if (k1.type > k2->type)
 210                 return 1;
 211         if (k1.type < k2->type)
 212                 return -1;
 213         if (k1.offset > k2->offset)
 214                 return 1;
 215         if (k1.offset < k2->offset)
 216                 return -1;
 217         return 0;
 218 }
 219
 220
 221 int btrfs_realloc_node(struct btrfs_trans_handle *trans,
 222                        struct btrfs_root *root, struct extent_buffer *parent,
 223                        int start_slot, int cache_only, u64 *last_ret,
 224                        struct btrfs_key *progress)
 225 {
 226         struct extent_buffer *cur;
 227         struct extent_buffer *tmp;
 228         u64 blocknr;
 229         u64 search_start = *last_ret;
 230         u64 last_block = 0;
 231         u64 other;
 232         u32 parent_nritems;
 233         int end_slot;
 234         int i;
 235         int err = 0;
 236         int parent_level;
 237         int uptodate;
 238         u32 blocksize;
 239         int progress_passed = 0;
 240         struct btrfs_disk_key disk_key;
 241
 242         parent_level = btrfs_header_level(parent);
 243         if (cache_only && parent_level != 1)
 244                 return 0;
 245
 246         if (trans->transaction != root->fs_info->running_transaction) {
 247                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 248                        root->fs_info->running_transaction->transid);
 249                 WARN_ON(1);
 250         }
 251         if (trans->transid != root->fs_info->generation) {
 252                 printk(KERN_CRIT "trans %Lu running %Lu\n", trans->transid,
 253                        root->fs_info->generation);
 254                 WARN_ON(1);
 255         }
 256
 257         parent_nritems = btrfs_header_nritems(parent);
 258         blocksize = btrfs_level_size(root, parent_level - 1);
 259         end_slot = parent_nritems;
 260
 261         if (parent_nritems == 1)
 262                 return 0;
 263
 264         for (i = start_slot; i < end_slot; i++) {
 265                 int close = 1;
 266
 267                 if (!parent->map_token) {
 268                         map_extent_buffer(parent,
 269                                         btrfs_node_key_ptr_offset(i),
 270                                         sizeof(struct btrfs_key_ptr),
 271                                         &parent->map_token, &parent->kaddr,
 272                                         &parent->map_start, &parent->map_len,
 273                                         KM_USER1);
 274                 }
 275                 btrfs_node_key(parent, &disk_key, i);
 276                 if (!progress_passed && comp_keys(&disk_key, progress) < 0)
 277                         continue;
 278
 279                 progress_passed = 1;
 280                 blocknr = btrfs_node_blockptr(parent, i);
 281                 if (last_block == 0)
 282                         last_block = blocknr;
 283
 284                 if (i > 0) {
 285                         other = btrfs_node_blockptr(parent, i - 1);
 286                         close = close_blocks(blocknr, other, blocksize);
 287                 }
 288                 if (close && i < end_slot - 2) {
 289                         other = btrfs_node_blockptr(parent, i + 1);
 290                         close = close_blocks(blocknr, other, blocksize);
 291                 }
 292                 if (close) {
 293                         last_block = blocknr;
 294                         continue;
 295                 }
 296                 if (parent->map_token) {
 297                         unmap_extent_buffer(parent, parent->map_token,
 298                                             KM_USER1);
 299                         parent->map_token = NULL;
 300                 }
 301
 302                 cur = btrfs_find_tree_block(root, blocknr, blocksize);
 303                 if (cur)
 304                         uptodate = btrfs_buffer_uptodate(cur);
 305                 else
 306                         uptodate = 0;
 307                 if (!cur || !uptodate) {
 308                         if (cache_only) {
 309                                 free_extent_buffer(cur);
 310                                 continue;
 311                         }
 312                         if (!cur) {
 313                                 cur = read_tree_block(root, blocknr,
 314                                                          blocksize);
 315                         } else if (!uptodate) {
 316                                 btrfs_read_buffer(cur);
 317                         }
 318                 }
 319                 if (search_start == 0)
 320                         search_start = last_block;
 321
 322                 err = __btrfs_cow_block(trans, root, cur, parent, i,
 323                                         &tmp, search_start,
 324                                         min(16 * blocksize,
 325                                             (end_slot - i) * blocksize));
 326                 if (err) {
 327                         free_extent_buffer(cur);
 328                         break;
 329                 }
 330                 search_start = tmp->start;
 331                 last_block = tmp->start;
 332                 *last_ret = search_start;
 333                 if (parent_level == 1)
 334                         btrfs_clear_buffer_defrag(tmp);
 335                 free_extent_buffer(tmp);
 336         }
 337         if (parent->map_token) {
 338                 unmap_extent_buffer(parent, parent->map_token,
 339                                     KM_USER1);
 340                 parent->map_token = NULL;
 341         }
 342         return err;
 343 }
 344
 345 /*
 346  * The leaf data grows from end-to-front in the node.
 347  * this returns the address of the start of the last item,
 348  * which is the stop of the leaf data stack
 349  */
 350 static inline unsigned int leaf_data_end(struct btrfs_root *root,
 351                                          struct extent_buffer *leaf)
 352 {
 353         u32 nr = btrfs_header_nritems(leaf);
 354         if (nr == 0)
 355                 return BTRFS_LEAF_DATA_SIZE(root);
 356         return btrfs_item_offset_nr(leaf, nr - 1);
 357 }
 358
 359 static int check_node(struct btrfs_root *root, struct btrfs_path *path,
 360                       int level)
 361 {
 362         struct extent_buffer *parent = NULL;
 363         struct extent_buffer *node = path->nodes[level];
 364         struct btrfs_disk_key parent_key;
 365         struct btrfs_disk_key node_key;
 366         int parent_slot;
 367         int slot;
 368         struct btrfs_key cpukey;
 369         u32 nritems = btrfs_header_nritems(node);
 370
 371         if (path->nodes[level + 1])
 372                 parent = path->nodes[level + 1];
 373
 374         slot = path->slots[level];
 375         BUG_ON(nritems == 0);
 376         if (parent) {
 377                 parent_slot = path->slots[level + 1];
 378                 btrfs_node_key(parent, &parent_key, parent_slot);
 379                 btrfs_node_key(node, &node_key, 0);
 380                 BUG_ON(memcmp(&parent_key, &node_key,
 381                               sizeof(struct btrfs_disk_key)));
 382                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 383                        btrfs_header_bytenr(node));
 384         }
 385         BUG_ON(nritems > BTRFS_NODEPTRS_PER_BLOCK(root));
 386         if (slot != 0) {
 387                 btrfs_node_key_to_cpu(node, &cpukey, slot - 1);
 388                 btrfs_node_key(node, &node_key, slot);
 389                 BUG_ON(comp_keys(&node_key, &cpukey) <= 0);
 390         }
 391         if (slot < nritems - 1) {
 392                 btrfs_node_key_to_cpu(node, &cpukey, slot + 1);
 393                 btrfs_node_key(node, &node_key, slot);
 394                 BUG_ON(comp_keys(&node_key, &cpukey) >= 0);
 395         }
 396         return 0;
 397 }
 398
 399 static int check_leaf(struct btrfs_root *root, struct btrfs_path *path,
 400                       int level)
 401 {
 402         struct extent_buffer *leaf = path->nodes[level];
 403         struct extent_buffer *parent = NULL;
 404         int parent_slot;
 405         struct btrfs_key cpukey;
 406         struct btrfs_disk_key parent_key;
 407         struct btrfs_disk_key leaf_key;
 408         int slot = path->slots[0];
 409
 410         u32 nritems = btrfs_header_nritems(leaf);
 411
 412         if (path->nodes[level + 1])
 413                 parent = path->nodes[level + 1];
 414
 415         if (nritems == 0)
 416                 return 0;
 417
 418         if (parent) {
 419                 parent_slot = path->slots[level + 1];
 420                 btrfs_node_key(parent, &parent_key, parent_slot);
 421                 btrfs_item_key(leaf, &leaf_key, 0);
 422
 423                 BUG_ON(memcmp(&parent_key, &leaf_key,
 424                        sizeof(struct btrfs_disk_key)));
 425                 BUG_ON(btrfs_node_blockptr(parent, parent_slot) !=
 426                        btrfs_header_bytenr(leaf));
 427         }
 428 #if 0
 429         for (i = 0; nritems > 1 && i < nritems - 2; i++) {
 430                 btrfs_item_key_to_cpu(leaf, &cpukey, i + 1);
 431                 btrfs_item_key(leaf, &leaf_key, i);
 432                 if (comp_keys(&leaf_key, &cpukey) >= 0) {
 433                         btrfs_print_leaf(root, leaf);
 434                         printk("slot %d offset bad key\n", i);
 435                         BUG_ON(1);
 436                 }
 437                 if (btrfs_item_offset_nr(leaf, i) !=
 438                         btrfs_item_end_nr(leaf, i + 1)) {
 439                         btrfs_print_leaf(root, leaf);
 440                         printk("slot %d offset bad\n", i);
 441                         BUG_ON(1);
 442                 }
 443                 if (i == 0) {
 444                         if (btrfs_item_offset_nr(leaf, i) +
 445                                btrfs_item_size_nr(leaf, i) !=
 446                                BTRFS_LEAF_DATA_SIZE(root)) {
 447                                 btrfs_print_leaf(root, leaf);
 448                                 printk("slot %d first offset bad\n", i);
 449                                 BUG_ON(1);
 450                         }
 451                 }
 452         }
 453         if (nritems > 0) {
 454                 if (btrfs_item_size_nr(leaf, nritems - 1) > 4096) {
 455                                 btrfs_print_leaf(root, leaf);
 456                                 printk("slot %d bad size \n", nritems - 1);
 457                                 BUG_ON(1);
 458                 }
 459         }
 460 #endif
 461         if (slot != 0 && slot < nritems - 1) {
 462                 btrfs_item_key(leaf, &leaf_key, slot);
 463                 btrfs_item_key_to_cpu(leaf, &cpukey, slot - 1);
 464                 if (comp_keys(&leaf_key, &cpukey) <= 0) {
 465                         btrfs_print_leaf(root, leaf);
 466                         printk("slot %d offset bad key\n", slot);
 467                         BUG_ON(1);
 468                 }
 469                 if (btrfs_item_offset_nr(leaf, slot - 1) !=
 470                        btrfs_item_end_nr(leaf, slot)) {
 471                         btrfs_print_leaf(root, leaf);
 472                         printk("slot %d offset bad\n", slot);
 473                         BUG_ON(1);
 474                 }
 475         }
 476         if (slot < nritems - 1) {
 477                 btrfs_item_key(leaf, &leaf_key, slot);
 478                 btrfs_item_key_to_cpu(leaf, &cpukey, slot + 1);
 479                 BUG_ON(comp_keys(&leaf_key, &cpukey) >= 0);
 480                 if (btrfs_item_offset_nr(leaf, slot) !=
 481                         btrfs_item_end_nr(leaf, slot + 1)) {
 482                         btrfs_print_leaf(root, leaf);
 483                         printk("slot %d offset bad\n", slot);
 484                         BUG_ON(1);
 485                 }
 486         }
 487         BUG_ON(btrfs_item_offset_nr(leaf, 0) +
 488                btrfs_item_size_nr(leaf, 0) != BTRFS_LEAF_DATA_SIZE(root));
 489         return 0;
 490 }
 491
 492 static int check_block(struct btrfs_root *root, struct btrfs_path *path,
 493                         int level)
 494 {
 495         return 0;
 496 #if 0
 497         struct extent_buffer *buf = path->nodes[level];
 498
 499         if (memcmp_extent_buffer(buf, root->fs_info->fsid,
 500                                  (unsigned long)btrfs_header_fsid(buf),
 501                                  BTRFS_FSID_SIZE)) {
 502                 printk("warning bad block %Lu\n", buf->start);
 503                 return 1;
 504         }
 505 #endif
 506         if (level == 0)
 507                 return check_leaf(root, path, level);
 508         return check_node(root, path, level);
 509 }
 510
 511 /*
 512  * search for key in the extent_buffer.  The items start at offset p,
 513  * and they are item_size apart.  There are 'max' items in p.
 514  *
 515  * the slot in the array is returned via slot, and it points to
 516  * the place where you would insert key if it is not found in
 517  * the array.
 518  *
 519  * slot may point to max if the key is bigger than all of the keys
 520  */
 521 static int generic_bin_search(struct extent_buffer *eb, unsigned long p,
 522                               int item_size, struct btrfs_key *key,
 523                               int max, int *slot)
 524 {
 525         int low = 0;
 526         int high = max;
 527         int mid;
 528         int ret;
 529         struct btrfs_disk_key *tmp = NULL;
 530         struct btrfs_disk_key unaligned;
 531         unsigned long offset;
 532         char *map_token = NULL;
 533         char *kaddr = NULL;
 534         unsigned long map_start = 0;
 535         unsigned long map_len = 0;
 536         int err;
 537
 538         while(low < high) {
 539                 mid = (low + high) / 2;
 540                 offset = p + mid * item_size;
 541
 542                 if (!map_token || offset < map_start ||
 543                     (offset + sizeof(struct btrfs_disk_key)) >
 544                     map_start + map_len) {
 545                         if (map_token) {
 546                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 547                                 map_token = NULL;
 548                         }
 549                         err = map_extent_buffer(eb, offset,
 550                                                 sizeof(struct btrfs_disk_key),
 551                                                 &map_token, &kaddr,
 552                                                 &map_start, &map_len, KM_USER0);
 553
 554                         if (!err) {
 555                                 tmp = (struct btrfs_disk_key *)(kaddr + offset -
 556                                                         map_start);
 557                         } else {
 558                                 read_extent_buffer(eb, &unaligned,
 559                                                    offset, sizeof(unaligned));
 560                                 tmp = &unaligned;
 561                         }
 562
 563                 } else {
 564                         tmp = (struct btrfs_disk_key *)(kaddr + offset -
 565                                                         map_start);
 566                 }
 567                 ret = comp_keys(tmp, key);
 568
 569                 if (ret < 0)
 570                         low = mid + 1;
 571                 else if (ret > 0)
 572                         high = mid;
 573                 else {
 574                         *slot = mid;
 575                         if (map_token)
 576                                 unmap_extent_buffer(eb, map_token, KM_USER0);
 577                         return 0;
 578                 }
 579         }
 580         *slot = low;
 581         if (map_token)
 582                 unmap_extent_buffer(eb, map_token, KM_USER0);
 583         return 1;
 584 }
 585
 586 /*
 587  * simple bin_search frontend that does the right thing for
 588  * leaves vs nodes
 589  */
 590 static int bin_search(struct extent_buffer *eb, struct btrfs_key *key,
 591                       int level, int *slot)
 592 {
 593         if (level == 0) {
 594                 return generic_bin_search(eb,
 595                                           offsetof(struct btrfs_leaf, items),
 596                                           sizeof(struct btrfs_item),
 597                                           key, btrfs_header_nritems(eb),
 598                                           slot);
 599         } else {
 600                 return generic_bin_search(eb,
 601                                           offsetof(struct btrfs_node, ptrs),
 602                                           sizeof(struct btrfs_key_ptr),
 603                                           key, btrfs_header_nritems(eb),
 604                                           slot);
 605         }
 606         return -1;
 607 }
 608
 609 static struct extent_buffer *read_node_slot(struct btrfs_root *root,
 610                                    struct extent_buffer *parent, int slot)
 611 {
 612         if (slot < 0)
 613                 return NULL;
 614         if (slot >= btrfs_header_nritems(parent))
 615                 return NULL;
 616         return read_tree_block(root, btrfs_node_blockptr(parent, slot),
 617                        btrfs_level_size(root, btrfs_header_level(parent) - 1));
 618 }
 619
 620 static int balance_level(struct btrfs_trans_handle *trans, struct btrfs_root
 621                          *root, struct btrfs_path *path, int level)
 622 {
 623         struct extent_buffer *right = NULL;
 624         struct extent_buffer *mid;
 625         struct extent_buffer *left = NULL;
 626         struct extent_buffer *parent = NULL;
 627         int ret = 0;
 628         int wret;
 629         int pslot;
 630         int orig_slot = path->slots[level];
 631         int err_on_enospc = 0;
 632         u64 orig_ptr;
 633
 634         if (level == 0)
 635                 return 0;
 636
 637         mid = path->nodes[level];
 638         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 639
 640         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 641
 642         if (level < BTRFS_MAX_LEVEL - 1)
 643                 parent = path->nodes[level + 1];
 644         pslot = path->slots[level + 1];
 645
 646         /*
 647          * deal with the case where there is only one pointer in the root
 648          * by promoting the node below to a root
 649          */
 650         if (!parent) {
 651                 struct extent_buffer *child;
 652
 653                 if (btrfs_header_nritems(mid) != 1)
 654                         return 0;
 655
 656                 /* promote the child to a root */
 657                 child = read_node_slot(root, mid, 0);
 658                 BUG_ON(!child);
 659                 root->node = child;
 660                 path->nodes[level] = NULL;
 661                 clean_tree_block(trans, root, mid);
 662                 wait_on_tree_block_writeback(root, mid);
 663                 /* once for the path */
 664                 free_extent_buffer(mid);
 665                 ret = btrfs_free_extent(trans, root, mid->start, mid->len,
 666                                         root->root_key.objectid,
 667                                         btrfs_header_generation(mid), 0, 0, 1);
 668                 /* once for the root ptr */
 669                 free_extent_buffer(mid);
 670                 return ret;
 671         }
 672         if (btrfs_header_nritems(mid) >
 673             BTRFS_NODEPTRS_PER_BLOCK(root) / 4)
 674                 return 0;
 675
 676         if (btrfs_header_nritems(mid) < 2)
 677                 err_on_enospc = 1;
 678
 679         left = read_node_slot(root, parent, pslot - 1);
 680         if (left) {
 681                 wret = btrfs_cow_block(trans, root, left,
 682                                        parent, pslot - 1, &left);
 683                 if (wret) {
 684                         ret = wret;
 685                         goto enospc;
 686                 }
 687         }
 688         right = read_node_slot(root, parent, pslot + 1);
 689         if (right) {
 690                 wret = btrfs_cow_block(trans, root, right,
 691                                        parent, pslot + 1, &right);
 692                 if (wret) {
 693                         ret = wret;
 694                         goto enospc;
 695                 }
 696         }
 697
 698         /* first, try to make some room in the middle buffer */
 699         if (left) {
 700                 orig_slot += btrfs_header_nritems(left);
 701                 wret = push_node_left(trans, root, left, mid);
 702                 if (wret < 0)
 703                         ret = wret;
 704                 if (btrfs_header_nritems(mid) < 2)
 705                         err_on_enospc = 1;
 706         }
 707
 708         /*
 709          * then try to empty the right most buffer into the middle
 710          */
 711         if (right) {
 712                 wret = push_node_left(trans, root, mid, right);
 713                 if (wret < 0 && wret != -ENOSPC)
 714                         ret = wret;
 715                 if (btrfs_header_nritems(right) == 0) {
 716                         u64 bytenr = right->start;
 717                         u64 generation = btrfs_header_generation(parent);
 718                         u32 blocksize = right->len;
 719
 720                         clean_tree_block(trans, root, right);
 721                         wait_on_tree_block_writeback(root, right);
 722                         free_extent_buffer(right);
 723                         right = NULL;
 724                         wret = del_ptr(trans, root, path, level + 1, pslot +
 725                                        1);
 726                         if (wret)
 727                                 ret = wret;
 728                         wret = btrfs_free_extent(trans, root, bytenr,
 729                                                  blocksize,
 730                                                  btrfs_header_owner(parent),
 731                                                  generation, 0, 0, 1);
 732                         if (wret)
 733                                 ret = wret;
 734                 } else {
 735                         struct btrfs_disk_key right_key;
 736                         btrfs_node_key(right, &right_key, 0);
 737                         btrfs_set_node_key(parent, &right_key, pslot + 1);
 738                         btrfs_mark_buffer_dirty(parent);
 739                 }
 740         }
 741         if (btrfs_header_nritems(mid) == 1) {
 742                 /*
 743                  * we're not allowed to leave a node with one item in the
 744                  * tree during a delete.  A deletion from lower in the tree
 745                  * could try to delete the only pointer in this node.
 746                  * So, pull some keys from the left.
 747                  * There has to be a left pointer at this point because
 748                  * otherwise we would have pulled some pointers from the
 749                  * right
 750                  */
 751                 BUG_ON(!left);
 752                 wret = balance_node_right(trans, root, mid, left);
 753                 if (wret < 0) {
 754                         ret = wret;
 755                         goto enospc;
 756                 }
 757                 BUG_ON(wret == 1);
 758         }
 759         if (btrfs_header_nritems(mid) == 0) {
 760                 /* we've managed to empty the middle node, drop it */
 761                 u64 root_gen = btrfs_header_generation(parent);
 762                 u64 bytenr = mid->start;
 763                 u32 blocksize = mid->len;
 764                 clean_tree_block(trans, root, mid);
 765                 wait_on_tree_block_writeback(root, mid);
 766                 free_extent_buffer(mid);
 767                 mid = NULL;
 768                 wret = del_ptr(trans, root, path, level + 1, pslot);
 769                 if (wret)
 770                         ret = wret;
 771                 wret = btrfs_free_extent(trans, root, bytenr, blocksize,
 772                                          btrfs_header_owner(parent),
 773                                          root_gen, 0, 0, 1);
 774                 if (wret)
 775                         ret = wret;
 776         } else {
 777                 /* update the parent key to reflect our changes */
 778                 struct btrfs_disk_key mid_key;
 779                 btrfs_node_key(mid, &mid_key, 0);
 780                 btrfs_set_node_key(parent, &mid_key, pslot);
 781                 btrfs_mark_buffer_dirty(parent);
 782         }
 783
 784         /* update the path */
 785         if (left) {
 786                 if (btrfs_header_nritems(left) > orig_slot) {
 787                         extent_buffer_get(left);
 788                         path->nodes[level] = left;
 789                         path->slots[level + 1] -= 1;
 790                         path->slots[level] = orig_slot;
 791                         if (mid)
 792                                 free_extent_buffer(mid);
 793                 } else {
 794                         orig_slot -= btrfs_header_nritems(left);
 795                         path->slots[level] = orig_slot;
 796                 }
 797         }
 798         /* double check we haven't messed things up */
 799         check_block(root, path, level);
 800         if (orig_ptr !=
 801             btrfs_node_blockptr(path->nodes[level], path->slots[level]))
 802                 BUG();
 803 enospc:
 804         if (right)
 805                 free_extent_buffer(right);
 806         if (left)
 807                 free_extent_buffer(left);
 808         return ret;
 809 }
 810
 811 /* returns zero if the push worked, non-zero otherwise */
 812 static int push_nodes_for_insert(struct btrfs_trans_handle *trans,
 813                                 struct btrfs_root *root,
 814                                 struct btrfs_path *path, int level)
 815 {
 816         struct extent_buffer *right = NULL;
 817         struct extent_buffer *mid;
 818         struct extent_buffer *left = NULL;
 819         struct extent_buffer *parent = NULL;
 820         int ret = 0;
 821         int wret;
 822         int pslot;
 823         int orig_slot = path->slots[level];
 824         u64 orig_ptr;
 825
 826         if (level == 0)
 827                 return 1;
 828
 829         mid = path->nodes[level];
 830         WARN_ON(btrfs_header_generation(mid) != trans->transid);
 831         orig_ptr = btrfs_node_blockptr(mid, orig_slot);
 832
 833         if (level < BTRFS_MAX_LEVEL - 1)
 834                 parent = path->nodes[level + 1];
 835         pslot = path->slots[level + 1];
 836
 837         if (!parent)
 838                 return 1;
 839
 840         left = read_node_slot(root, parent, pslot - 1);
 841
 842         /* first, try to make some room in the middle buffer */
 843         if (left) {
 844                 u32 left_nr;
 845                 left_nr = btrfs_header_nritems(left);
 846                 if (left_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 847                         wret = 1;
 848                 } else {
 849                         ret = btrfs_cow_block(trans, root, left, parent,
 850                                               pslot - 1, &left);
 851                         if (ret)
 852                                 wret = 1;
 853                         else {
 854                                 wret = push_node_left(trans, root,
 855                                                       left, mid);
 856                         }
 857                 }
 858                 if (wret < 0)
 859                         ret = wret;
 860                 if (wret == 0) {
 861                         struct btrfs_disk_key disk_key;
 862                         orig_slot += left_nr;
 863                         btrfs_node_key(mid, &disk_key, 0);
 864                         btrfs_set_node_key(parent, &disk_key, pslot);
 865                         btrfs_mark_buffer_dirty(parent);
 866                         if (btrfs_header_nritems(left) > orig_slot) {
 867                                 path->nodes[level] = left;
 868                                 path->slots[level + 1] -= 1;
 869                                 path->slots[level] = orig_slot;
 870                                 free_extent_buffer(mid);
 871                         } else {
 872                                 orig_slot -=
 873                                         btrfs_header_nritems(left);
 874                                 path->slots[level] = orig_slot;
 875                                 free_extent_buffer(left);
 876                         }
 877                         return 0;
 878                 }
 879                 free_extent_buffer(left);
 880         }
 881         right= read_node_slot(root, parent, pslot + 1);
 882
 883         /*
 884          * then try to empty the right most buffer into the middle
 885          */
 886         if (right) {
 887                 u32 right_nr;
 888                 right_nr = btrfs_header_nritems(right);
 889                 if (right_nr >= BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
 890                         wret = 1;
 891                 } else {
 892                         ret = btrfs_cow_block(trans, root, right,
 893                                               parent, pslot + 1,
 894                                               &right);
 895                         if (ret)
 896                                 wret = 1;
 897                         else {
 898                                 wret = balance_node_right(trans, root,
 899                                                           right, mid);
 900                         }
 901                 }
 902                 if (wret < 0)
 903                         ret = wret;
 904                 if (wret == 0) {
 905                         struct btrfs_disk_key disk_key;
 906
 907                         btrfs_node_key(right, &disk_key, 0);
 908                         btrfs_set_node_key(parent, &disk_key, pslot + 1);
 909                         btrfs_mark_buffer_dirty(parent);
 910
 911                         if (btrfs_header_nritems(mid) <= orig_slot) {
 912                                 path->nodes[level] = right;
 913                                 path->slots[level + 1] += 1;
 914                                 path->slots[level] = orig_slot -
 915                                         btrfs_header_nritems(mid);
 916                                 free_extent_buffer(mid);
 917                         } else {
 918                                 free_extent_buffer(right);
 919                         }
 920                         return 0;
 921                 }
 922                 free_extent_buffer(right);
 923         }
 924         return 1;
 925 }
 926
 927 /*
 928  * readahead one full node of leaves
 929  */
 930 static void reada_for_search(struct btrfs_root *root, struct btrfs_path *path,
 931                              int level, int slot)
 932 {
 933         struct extent_buffer *node;
 934         u32 nritems;
 935         u64 search;
 936         u64 lowest_read;
 937         u64 highest_read;
 938         u64 nread = 0;
 939         int direction = path->reada;
 940         struct extent_buffer *eb;
 941         u32 nr;
 942         u32 blocksize;
 943         u32 nscan = 0;
 944
 945         if (level != 1)
 946                 return;
 947
 948         if (!path->nodes[level])
 949                 return;
 950
 951         node = path->nodes[level];
 952         search = btrfs_node_blockptr(node, slot);
 953         blocksize = btrfs_level_size(root, level - 1);
 954         eb = btrfs_find_tree_block(root, search, blocksize);
 955         if (eb) {
 956                 free_extent_buffer(eb);
 957                 return;
 958         }
 959
 960         highest_read = search;
 961         lowest_read = search;
 962
 963         nritems = btrfs_header_nritems(node);
 964         nr = slot;
 965         while(1) {
 966                 if (direction < 0) {
 967                         if (nr == 0)
 968                                 break;
 969                         nr--;
 970                 } else if (direction > 0) {
 971                         nr++;
 972                         if (nr >= nritems)
 973                                 break;
 974                 }
 975                 search = btrfs_node_blockptr(node, nr);
 976                 if ((search >= lowest_read && search <= highest_read) ||
 977                     (search < lowest_read && lowest_read - search <= 32768) ||
 978                     (search > highest_read && search - highest_read <= 32768)) {
 979                         readahead_tree_block(root, search, blocksize);
 980                         nread += blocksize;
 981                 }
 982                 nscan++;
 983                 if (path->reada < 2 && (nread > (256 * 1024) || nscan > 32))
 984                         break;
 985                 if(nread > (1024 * 1024) || nscan > 128)
 986                         break;
 987
 988                 if (search < lowest_read)
 989                         lowest_read = search;
 990                 if (search > highest_read)
 991                         highest_read = search;
 992         }
 993 }
 994 /*
 995  * look for key in the tree.  path is filled in with nodes along the way
 996  * if key is found, we return zero and you can find the item in the leaf
 997  * level of the path (level 0)
 998  *
 999  * If the key isn't found, the path points to the slot where it should
1000  * be inserted, and 1 is returned.  If there are other errors during the
1001  * search a negative error number is returned.
1002  *
1003  * if ins_len > 0, nodes and leaves will be split as we walk down the
1004  * tree.  if ins_len < 0, nodes will be merged as we walk down the tree (if
1005  * possible)
1006  */
1007 int btrfs_search_slot(struct btrfs_trans_handle *trans, struct btrfs_root
1008                       *root, struct btrfs_key *key, struct btrfs_path *p, int
1009                       ins_len, int cow)
1010 {
1011         struct extent_buffer *b;
1012         u64 bytenr;
1013         u64 ptr_gen;
1014         int slot;
1015         int ret;
1016         int level;
1017         int should_reada = p->reada;
1018         u8 lowest_level = 0;
1019
1020         lowest_level = p->lowest_level;
1021         WARN_ON(lowest_level && ins_len);
1022         WARN_ON(p->nodes[0] != NULL);
1023         WARN_ON(!mutex_is_locked(&root->fs_info->fs_mutex));
1024 again:
1025         b = root->node;
1026         extent_buffer_get(b);
1027         while (b) {
1028                 level = btrfs_header_level(b);
1029                 if (cow) {
1030                         int wret;
1031                         wret = btrfs_cow_block(trans, root, b,
1032                                                p->nodes[level + 1],
1033                                                p->slots[level + 1],
1034                                                &b);
1035                         if (wret) {
1036                                 free_extent_buffer(b);
1037                                 return wret;
1038                         }
1039                 }
1040                 BUG_ON(!cow && ins_len);
1041                 if (level != btrfs_header_level(b))
1042                         WARN_ON(1);
1043                 level = btrfs_header_level(b);
1044                 p->nodes[level] = b;
1045                 ret = check_block(root, p, level);
1046                 if (ret)
1047                         return -1;
1048                 ret = bin_search(b, key, level, &slot);
1049                 if (level != 0) {
1050                         if (ret && slot > 0)
1051                                 slot -= 1;
1052                         p->slots[level] = slot;
1053                         if (ins_len > 0 && btrfs_header_nritems(b) >=
1054                             BTRFS_NODEPTRS_PER_BLOCK(root) - 1) {
1055                                 int sret = split_node(trans, root, p, level);
1056                                 BUG_ON(sret > 0);
1057                                 if (sret)
1058                                         return sret;
1059                                 b = p->nodes[level];
1060                                 slot = p->slots[level];
1061                         } else if (ins_len < 0) {
1062                                 int sret = balance_level(trans, root, p,
1063                                                          level);
1064                                 if (sret)
1065                                         return sret;
1066                                 b = p->nodes[level];
1067                                 if (!b) {
1068                                         btrfs_release_path(NULL, p);
1069                                         goto again;
1070                                 }
1071                                 slot = p->slots[level];
1072                                 BUG_ON(btrfs_header_nritems(b) == 1);
1073                         }
1074                         /* this is only true while dropping a snapshot */
1075                         if (level == lowest_level)
1076                                 break;
1077                         bytenr = btrfs_node_blockptr(b, slot);
1078                         ptr_gen = btrfs_node_ptr_generation(b, slot);
1079                         if (should_reada)
1080                                 reada_for_search(root, p, level, slot);
1081                         b = read_tree_block(root, bytenr,
1082                                             btrfs_level_size(root, level - 1));
1083                         if (ptr_gen != btrfs_header_generation(b)) {
1084                                 printk("block %llu bad gen wanted %llu "
1085                                        "found %llu\n",
1086                                 (unsigned long long)b->start,
1087                                 (unsigned long long)ptr_gen,
1088                                 (unsigned long long)btrfs_header_generation(b));
1089                         }
1090                 } else {
1091                         p->slots[level] = slot;
1092                         if (ins_len > 0 && btrfs_leaf_free_space(root, b) <
1093                             sizeof(struct btrfs_item) + ins_len) {
1094                                 int sret = split_leaf(trans, root, key,
1095                                                       p, ins_len, ret == 0);
1096                                 BUG_ON(sret > 0);
1097                                 if (sret)
1098                                         return sret;
1099                         }
1100                         return ret;
1101                 }
1102         }
1103         return 1;
1104 }
1105
1106 /*
1107  * adjust the pointers going up the tree, starting at level
1108  * making sure the right key of each node is points to 'key'.
1109  * This is used after shifting pointers to the left, so it stops
1110  * fixing up pointers when a given leaf/node is not in slot 0 of the
1111  * higher levels
1112  *
1113  * If this fails to write a tree block, it returns -1, but continues
1114  * fixing up the blocks in ram so the tree is consistent.
1115  */
1116 static int fixup_low_keys(struct btrfs_trans_handle *trans,
1117                           struct btrfs_root *root, struct btrfs_path *path,
1118                           struct btrfs_disk_key *key, int level)
1119 {
1120         int i;
1121         int ret = 0;
1122         struct extent_buffer *t;
1123
1124         for (i = level; i < BTRFS_MAX_LEVEL; i++) {
1125                 int tslot = path->slots[i];
1126                 if (!path->nodes[i])
1127                         break;
1128                 t = path->nodes[i];
1129                 btrfs_set_node_key(t, key, tslot);
1130                 btrfs_mark_buffer_dirty(path->nodes[i]);
1131                 if (tslot != 0)
1132                         break;
1133         }
1134         return ret;
1135 }
1136
1137 /*
1138  * try to push data from one node into the next node left in the
1139  * tree.
1140  *
1141  * returns 0 if some ptrs were pushed left, < 0 if there was some horrible
1142  * error, and > 0 if there was no room in the left hand block.
1143  */
1144 static int push_node_left(struct btrfs_trans_handle *trans, struct btrfs_root
1145                           *root, struct extent_buffer *dst,
1146                           struct extent_buffer *src)
1147 {
1148         int push_items = 0;
1149         int src_nritems;
1150         int dst_nritems;
1151         int ret = 0;
1152
1153         src_nritems = btrfs_header_nritems(src);
1154         dst_nritems = btrfs_header_nritems(dst);
1155         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1156         WARN_ON(btrfs_header_generation(src) != trans->transid);
1157         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1158
1159         if (push_items <= 0) {
1160                 return 1;
1161         }
1162
1163         if (src_nritems < push_items)
1164                 push_items = src_nritems;
1165
1166         copy_extent_buffer(dst, src,
1167                            btrfs_node_key_ptr_offset(dst_nritems),
1168                            btrfs_node_key_ptr_offset(0),
1169                            push_items * sizeof(struct btrfs_key_ptr));
1170
1171         if (push_items < src_nritems) {
1172                 memmove_extent_buffer(src, btrfs_node_key_ptr_offset(0),
1173                                       btrfs_node_key_ptr_offset(push_items),
1174                                       (src_nritems - push_items) *
1175                                       sizeof(struct btrfs_key_ptr));
1176         }
1177         btrfs_set_header_nritems(src, src_nritems - push_items);
1178         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1179         btrfs_mark_buffer_dirty(src);
1180         btrfs_mark_buffer_dirty(dst);
1181         return ret;
1182 }
1183
1184 /*
1185  * try to push data from one node into the next node right in the
1186  * tree.
1187  *
1188  * returns 0 if some ptrs were pushed, < 0 if there was some horrible
1189  * error, and > 0 if there was no room in the right hand block.
1190  *
1191  * this will  only push up to 1/2 the contents of the left node over
1192  */
1193 static int balance_node_right(struct btrfs_trans_handle *trans,
1194                               struct btrfs_root *root,
1195                               struct extent_buffer *dst,
1196                               struct extent_buffer *src)
1197 {
1198         int push_items = 0;
1199         int max_push;
1200         int src_nritems;
1201         int dst_nritems;
1202         int ret = 0;
1203
1204         WARN_ON(btrfs_header_generation(src) != trans->transid);
1205         WARN_ON(btrfs_header_generation(dst) != trans->transid);
1206
1207         src_nritems = btrfs_header_nritems(src);
1208         dst_nritems = btrfs_header_nritems(dst);
1209         push_items = BTRFS_NODEPTRS_PER_BLOCK(root) - dst_nritems;
1210         if (push_items <= 0)
1211                 return 1;
1212
1213         max_push = src_nritems / 2 + 1;
1214         /* don't try to empty the node */
1215         if (max_push >= src_nritems)
1216                 return 1;
1217
1218         if (max_push < push_items)
1219                 push_items = max_push;
1220
1221         memmove_extent_buffer(dst, btrfs_node_key_ptr_offset(push_items),
1222                                       btrfs_node_key_ptr_offset(0),
1223                                       (dst_nritems) *
1224                                       sizeof(struct btrfs_key_ptr));
1225
1226         copy_extent_buffer(dst, src,
1227                            btrfs_node_key_ptr_offset(0),
1228                            btrfs_node_key_ptr_offset(src_nritems - push_items),
1229                            push_items * sizeof(struct btrfs_key_ptr));
1230
1231         btrfs_set_header_nritems(src, src_nritems - push_items);
1232         btrfs_set_header_nritems(dst, dst_nritems + push_items);
1233
1234         btrfs_mark_buffer_dirty(src);
1235         btrfs_mark_buffer_dirty(dst);
1236         return ret;
1237 }
1238
1239 /*
1240  * helper function to insert a new root level in the tree.
1241  * A new node is allocated, and a single item is inserted to
1242  * point to the existing root
1243  *
1244  * returns zero on success or < 0 on failure.
1245  */
1246 static int insert_new_root(struct btrfs_trans_handle *trans,
1247                            struct btrfs_root *root,
1248                            struct btrfs_path *path, int level)
1249 {
1250         u64 root_gen;
1251         u64 lower_gen;
1252         struct extent_buffer *lower;
1253         struct extent_buffer *c;
1254         struct btrfs_disk_key lower_key;
1255
1256         BUG_ON(path->nodes[level]);
1257         BUG_ON(path->nodes[level-1] != root->node);
1258
1259         if (root->ref_cows)
1260                 root_gen = trans->transid;
1261         else
1262                 root_gen = 0;
1263
1264         lower = path->nodes[level-1];
1265         if (level == 1)
1266                 btrfs_item_key(lower, &lower_key, 0);
1267         else
1268                 btrfs_node_key(lower, &lower_key, 0);
1269
1270         c = __btrfs_alloc_free_block(trans, root, root->nodesize,
1271                                    root->root_key.objectid,
1272                                    root_gen, lower_key.objectid, level,
1273                                    root->node->start, 0);
1274         if (IS_ERR(c))
1275                 return PTR_ERR(c);
1276         memset_extent_buffer(c, 0, 0, root->nodesize);
1277         btrfs_set_header_nritems(c, 1);
1278         btrfs_set_header_level(c, level);
1279         btrfs_set_header_bytenr(c, c->start);
1280         btrfs_set_header_generation(c, trans->transid);
1281         btrfs_set_header_owner(c, root->root_key.objectid);
1282
1283         write_extent_buffer(c, root->fs_info->fsid,
1284                             (unsigned long)btrfs_header_fsid(c),
1285                             BTRFS_FSID_SIZE);
1286         btrfs_set_node_key(c, &lower_key, 0);
1287         btrfs_set_node_blockptr(c, 0, lower->start);
1288         lower_gen = btrfs_header_generation(lower);
1289         WARN_ON(lower_gen == 0);
1290
1291         btrfs_set_node_ptr_generation(c, 0, lower_gen);
1292
1293         btrfs_mark_buffer_dirty(c);
1294
1295         /* the super has an extra ref to root->node */
1296         free_extent_buffer(root->node);
1297         root->node = c;
1298         extent_buffer_get(c);
1299         path->nodes[level] = c;
1300         path->slots[level] = 0;
1301
1302         if (root->ref_cows && lower_gen != trans->transid) {
1303                 struct btrfs_path *back_path = btrfs_alloc_path();
1304                 int ret;
1305                 ret = btrfs_insert_extent_backref(trans,
1306                                                   root->fs_info->extent_root,
1307                                                   path, lower->start,
1308                                                   root->root_key.objectid,
1309                                                   trans->transid, 0, 0);
1310                 BUG_ON(ret);
1311                 btrfs_free_path(back_path);
1312         }
1313         return 0;
1314 }
1315
1316 /*
1317  * worker function to insert a single pointer in a node.
1318  * the node should have enough room for the pointer already
1319  *
1320  * slot and level indicate where you want the key to go, and
1321  * blocknr is the block the key points to.
1322  *
1323  * returns zero on success and < 0 on any error
1324  */
1325 static int insert_ptr(struct btrfs_trans_handle *trans, struct btrfs_root
1326                       *root, struct btrfs_path *path, struct btrfs_disk_key
1327                       *key, u64 bytenr, int slot, int level)
1328 {
1329         struct extent_buffer *lower;
1330         int nritems;
1331
1332         BUG_ON(!path->nodes[level]);
1333         lower = path->nodes[level];
1334         nritems = btrfs_header_nritems(lower);
1335         if (slot > nritems)
1336                 BUG();
1337         if (nritems == BTRFS_NODEPTRS_PER_BLOCK(root))
1338                 BUG();
1339         if (slot != nritems) {
1340                 memmove_extent_buffer(lower,
1341                               btrfs_node_key_ptr_offset(slot + 1),
1342                               btrfs_node_key_ptr_offset(slot),
1343                               (nritems - slot) * sizeof(struct btrfs_key_ptr));
1344         }
1345         btrfs_set_node_key(lower, key, slot);
1346         btrfs_set_node_blockptr(lower, slot, bytenr);
1347         WARN_ON(trans->transid == 0);
1348         btrfs_set_node_ptr_generation(lower, slot, trans->transid);
1349         btrfs_set_header_nritems(lower, nritems + 1);
1350         btrfs_mark_buffer_dirty(lower);
1351         return 0;
1352 }
1353
1354 /*
1355  * split the node at the specified level in path in two.
1356  * The path is corrected to point to the appropriate node after the split
1357  *
1358  * Before splitting this tries to make some room in the node by pushing
1359  * left and right, if either one works, it returns right away.
1360  *
1361  * returns 0 on success and < 0 on failure
1362  */
1363 static int split_node(struct btrfs_trans_handle *trans, struct btrfs_root
1364                       *root, struct btrfs_path *path, int level)
1365 {
1366         u64 root_gen;
1367         struct extent_buffer *c;
1368         struct extent_buffer *split;
1369         struct btrfs_disk_key disk_key;
1370         int mid;
1371         int ret;
1372         int wret;
1373         u32 c_nritems;
1374
1375         c = path->nodes[level];
1376         WARN_ON(btrfs_header_generation(c) != trans->transid);
1377         if (c == root->node) {
1378                 /* trying to split the root, lets make a new one */
1379                 ret = insert_new_root(trans, root, path, level + 1);
1380                 if (ret)
1381                         return ret;
1382         } else {
1383                 ret = push_nodes_for_insert(trans, root, path, level);
1384                 c = path->nodes[level];
1385                 if (!ret && btrfs_header_nritems(c) <
1386                     BTRFS_NODEPTRS_PER_BLOCK(root) - 1)
1387                         return 0;
1388                 if (ret < 0)
1389                         return ret;
1390         }
1391
1392         c_nritems = btrfs_header_nritems(c);
1393         if (root->ref_cows)
1394                 root_gen = trans->transid;
1395         else
1396                 root_gen = 0;
1397
1398         btrfs_node_key(c, &disk_key, 0);
1399         split = __btrfs_alloc_free_block(trans, root, root->nodesize,
1400                                          root->root_key.objectid,
1401                                          root_gen,
1402                                          btrfs_disk_key_objectid(&disk_key),
1403                                          level, c->start, 0);
1404         if (IS_ERR(split))
1405                 return PTR_ERR(split);
1406
1407         btrfs_set_header_flags(split, btrfs_header_flags(c));
1408         btrfs_set_header_level(split, btrfs_header_level(c));
1409         btrfs_set_header_bytenr(split, split->start);
1410         btrfs_set_header_generation(split, trans->transid);
1411         btrfs_set_header_owner(split, root->root_key.objectid);
1412         write_extent_buffer(split, root->fs_info->fsid,
1413                             (unsigned long)btrfs_header_fsid(split),
1414                             BTRFS_FSID_SIZE);
1415
1416         mid = (c_nritems + 1) / 2;
1417
1418         copy_extent_buffer(split, c,
1419                            btrfs_node_key_ptr_offset(0),
1420                            btrfs_node_key_ptr_offset(mid),
1421                            (c_nritems - mid) * sizeof(struct btrfs_key_ptr));
1422         btrfs_set_header_nritems(split, c_nritems - mid);
1423         btrfs_set_header_nritems(c, mid);
1424         ret = 0;
1425
1426         btrfs_mark_buffer_dirty(c);
1427         btrfs_mark_buffer_dirty(split);
1428
1429         btrfs_node_key(split, &disk_key, 0);
1430         wret = insert_ptr(trans, root, path, &disk_key, split->start,
1431                           path->slots[level + 1] + 1,
1432                           level + 1);
1433         if (wret)
1434                 ret = wret;
1435
1436         if (path->slots[level] >= mid) {
1437                 path->slots[level] -= mid;
1438                 free_extent_buffer(c);
1439                 path->nodes[level] = split;
1440                 path->slots[level + 1] += 1;
1441         } else {
1442                 free_extent_buffer(split);
1443         }
1444         return ret;
1445 }
1446
1447 /*
1448  * how many bytes are required to store the items in a leaf.  start
1449  * and nr indicate which items in the leaf to check.  This totals up the
1450  * space used both by the item structs and the item data
1451  */
1452 static int leaf_space_used(struct extent_buffer *l, int start, int nr)
1453 {
1454         int data_len;
1455         int nritems = btrfs_header_nritems(l);
1456         int end = min(nritems, start + nr) - 1;
1457
1458         if (!nr)
1459                 return 0;
1460         data_len = btrfs_item_end_nr(l, start);
1461         data_len = data_len - btrfs_item_offset_nr(l, end);
1462         data_len += sizeof(struct btrfs_item) * nr;
1463         WARN_ON(data_len < 0);
1464         return data_len;
1465 }
1466
1467 /*
1468  * The space between the end of the leaf items and
1469  * the start of the leaf data.  IOW, how much room
1470  * the leaf has left for both items and data
1471  */
1472 int btrfs_leaf_free_space(struct btrfs_root *root, struct extent_buffer *leaf)
1473 {
1474         int nritems = btrfs_header_nritems(leaf);
1475         int ret;
1476         ret = BTRFS_LEAF_DATA_SIZE(root) - leaf_space_used(leaf, 0, nritems);
1477         if (ret < 0) {
1478                 printk("leaf free space ret %d, leaf data size %lu, used %d nritems %d\n",
1479                        ret, (unsigned long) BTRFS_LEAF_DATA_SIZE(root),
1480                        leaf_space_used(leaf, 0, nritems), nritems);
1481         }
1482         return ret;
1483 }
1484
1485 /*
1486  * push some data in the path leaf to the right, trying to free up at
1487  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1488  *
1489  * returns 1 if the push failed because the other node didn't have enough
1490  * room, 0 if everything worked out and < 0 if there were major errors.
1491  */
1492 static int push_leaf_right(struct btrfs_trans_handle *trans, struct btrfs_root
1493                            *root, struct btrfs_path *path, int data_size,
1494                            int empty)
1495 {
1496         struct extent_buffer *left = path->nodes[0];
1497         struct extent_buffer *right;
1498         struct extent_buffer *upper;
1499         struct btrfs_disk_key disk_key;
1500         int slot;
1501         u32 i;
1502         int free_space;
1503         int push_space = 0;
1504         int push_items = 0;
1505         struct btrfs_item *item;
1506         u32 left_nritems;
1507         u32 nr;
1508         u32 right_nritems;
1509         u32 data_end;
1510         u32 this_item_size;
1511         int ret;
1512
1513         slot = path->slots[1];
1514         if (!path->nodes[1]) {
1515                 return 1;
1516         }
1517         upper = path->nodes[1];
1518         if (slot >= btrfs_header_nritems(upper) - 1)
1519                 return 1;
1520
1521         right = read_tree_block(root, btrfs_node_blockptr(upper, slot + 1),
1522                                 root->leafsize);
1523         free_space = btrfs_leaf_free_space(root, right);
1524         if (free_space < data_size + sizeof(struct btrfs_item)) {
1525                 free_extent_buffer(right);
1526                 return 1;
1527         }
1528
1529         /* cow and double check */
1530         ret = btrfs_cow_block(trans, root, right, upper,
1531                               slot + 1, &right);
1532         if (ret) {
1533                 free_extent_buffer(right);
1534                 return 1;
1535         }
1536         free_space = btrfs_leaf_free_space(root, right);
1537         if (free_space < data_size + sizeof(struct btrfs_item)) {
1538                 free_extent_buffer(right);
1539                 return 1;
1540         }
1541
1542         left_nritems = btrfs_header_nritems(left);
1543         if (left_nritems == 0) {
1544                 free_extent_buffer(right);
1545                 return 1;
1546         }
1547
1548         if (empty)
1549                 nr = 0;
1550         else
1551                 nr = 1;
1552
1553         i = left_nritems - 1;
1554         while (i >= nr) {
1555                 item = btrfs_item_nr(left, i);
1556
1557                 if (path->slots[0] == i)
1558                         push_space += data_size + sizeof(*item);
1559
1560                 if (!left->map_token) {
1561                         map_extent_buffer(left, (unsigned long)item,
1562                                         sizeof(struct btrfs_item),
1563                                         &left->map_token, &left->kaddr,
1564                                         &left->map_start, &left->map_len,
1565                                         KM_USER1);
1566                 }
1567
1568                 this_item_size = btrfs_item_size(left, item);
1569                 if (this_item_size + sizeof(*item) + push_space > free_space)
1570                         break;
1571                 push_items++;
1572                 push_space += this_item_size + sizeof(*item);
1573                 if (i == 0)
1574                         break;
1575                 i--;
1576         }
1577         if (left->map_token) {
1578                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1579                 left->map_token = NULL;
1580         }
1581
1582         if (push_items == 0) {
1583                 free_extent_buffer(right);
1584                 return 1;
1585         }
1586
1587         if (!empty && push_items == left_nritems)
1588                 WARN_ON(1);
1589
1590         /* push left to right */
1591         right_nritems = btrfs_header_nritems(right);
1592
1593         push_space = btrfs_item_end_nr(left, left_nritems - push_items);
1594         push_space -= leaf_data_end(root, left);
1595
1596         /* make room in the right data area */
1597         data_end = leaf_data_end(root, right);
1598         memmove_extent_buffer(right,
1599                               btrfs_leaf_data(right) + data_end - push_space,
1600                               btrfs_leaf_data(right) + data_end,
1601                               BTRFS_LEAF_DATA_SIZE(root) - data_end);
1602
1603         /* copy from the left data area */
1604         copy_extent_buffer(right, left, btrfs_leaf_data(right) +
1605                      BTRFS_LEAF_DATA_SIZE(root) - push_space,
1606                      btrfs_leaf_data(left) + leaf_data_end(root, left),
1607                      push_space);
1608
1609         memmove_extent_buffer(right, btrfs_item_nr_offset(push_items),
1610                               btrfs_item_nr_offset(0),
1611                               right_nritems * sizeof(struct btrfs_item));
1612
1613         /* copy the items from left to right */
1614         copy_extent_buffer(right, left, btrfs_item_nr_offset(0),
1615                    btrfs_item_nr_offset(left_nritems - push_items),
1616                    push_items * sizeof(struct btrfs_item));
1617
1618         /* update the item pointers */
1619         right_nritems += push_items;
1620         btrfs_set_header_nritems(right, right_nritems);
1621         push_space = BTRFS_LEAF_DATA_SIZE(root);
1622         for (i = 0; i < right_nritems; i++) {
1623                 item = btrfs_item_nr(right, i);
1624                 if (!right->map_token) {
1625                         map_extent_buffer(right, (unsigned long)item,
1626                                         sizeof(struct btrfs_item),
1627                                         &right->map_token, &right->kaddr,
1628                                         &right->map_start, &right->map_len,
1629                                         KM_USER1);
1630                 }
1631                 push_space -= btrfs_item_size(right, item);
1632                 btrfs_set_item_offset(right, item, push_space);
1633         }
1634
1635         if (right->map_token) {
1636                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1637                 right->map_token = NULL;
1638         }
1639         left_nritems -= push_items;
1640         btrfs_set_header_nritems(left, left_nritems);
1641
1642         if (left_nritems)
1643                 btrfs_mark_buffer_dirty(left);
1644         btrfs_mark_buffer_dirty(right);
1645
1646         btrfs_item_key(right, &disk_key, 0);
1647         btrfs_set_node_key(upper, &disk_key, slot + 1);
1648         btrfs_mark_buffer_dirty(upper);
1649
1650         /* then fixup the leaf pointer in the path */
1651         if (path->slots[0] >= left_nritems) {
1652                 path->slots[0] -= left_nritems;
1653                 free_extent_buffer(path->nodes[0]);
1654                 path->nodes[0] = right;
1655                 path->slots[1] += 1;
1656         } else {
1657                 free_extent_buffer(right);
1658         }
1659         return 0;
1660 }
1661 /*
1662  * push some data in the path leaf to the left, trying to free up at
1663  * least data_size bytes.  returns zero if the push worked, nonzero otherwise
1664  */
1665 static int push_leaf_left(struct btrfs_trans_handle *trans, struct btrfs_root
1666                           *root, struct btrfs_path *path, int data_size,
1667                           int empty)
1668 {
1669         struct btrfs_disk_key disk_key;
1670         struct extent_buffer *right = path->nodes[0];
1671         struct extent_buffer *left;
1672         int slot;
1673         int i;
1674         int free_space;
1675         int push_space = 0;
1676         int push_items = 0;
1677         struct btrfs_item *item;
1678         u32 old_left_nritems;
1679         u32 right_nritems;
1680         u32 nr;
1681         int ret = 0;
1682         int wret;
1683         u32 this_item_size;
1684         u32 old_left_item_size;
1685
1686         slot = path->slots[1];
1687         if (slot == 0)
1688                 return 1;
1689         if (!path->nodes[1])
1690                 return 1;
1691
1692         right_nritems = btrfs_header_nritems(right);
1693         if (right_nritems == 0) {
1694                 return 1;
1695         }
1696
1697         left = read_tree_block(root, btrfs_node_blockptr(path->nodes[1],
1698                                slot - 1), root->leafsize);
1699         free_space = btrfs_leaf_free_space(root, left);
1700         if (free_space < data_size + sizeof(struct btrfs_item)) {
1701                 free_extent_buffer(left);
1702                 return 1;
1703         }
1704
1705         /* cow and double check */
1706         ret = btrfs_cow_block(trans, root, left,
1707                               path->nodes[1], slot - 1, &left);
1708         if (ret) {
1709                 /* we hit -ENOSPC, but it isn't fatal here */
1710                 free_extent_buffer(left);
1711                 return 1;
1712         }
1713
1714         free_space = btrfs_leaf_free_space(root, left);
1715         if (free_space < data_size + sizeof(struct btrfs_item)) {
1716                 free_extent_buffer(left);
1717                 return 1;
1718         }
1719
1720         if (empty)
1721                 nr = right_nritems;
1722         else
1723                 nr = right_nritems - 1;
1724
1725         for (i = 0; i < nr; i++) {
1726                 item = btrfs_item_nr(right, i);
1727                 if (!right->map_token) {
1728                         map_extent_buffer(right, (unsigned long)item,
1729                                         sizeof(struct btrfs_item),
1730                                         &right->map_token, &right->kaddr,
1731                                         &right->map_start, &right->map_len,
1732                                         KM_USER1);
1733                 }
1734
1735                 if (path->slots[0] == i)
1736                         push_space += data_size + sizeof(*item);
1737
1738                 this_item_size = btrfs_item_size(right, item);
1739                 if (this_item_size + sizeof(*item) + push_space > free_space)
1740                         break;
1741
1742                 push_items++;
1743                 push_space += this_item_size + sizeof(*item);
1744         }
1745
1746         if (right->map_token) {
1747                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1748                 right->map_token = NULL;
1749         }
1750
1751         if (push_items == 0) {
1752                 free_extent_buffer(left);
1753                 return 1;
1754         }
1755         if (!empty && push_items == btrfs_header_nritems(right))
1756                 WARN_ON(1);
1757
1758         /* push data from right to left */
1759         copy_extent_buffer(left, right,
1760                            btrfs_item_nr_offset(btrfs_header_nritems(left)),
1761                            btrfs_item_nr_offset(0),
1762                            push_items * sizeof(struct btrfs_item));
1763
1764         push_space = BTRFS_LEAF_DATA_SIZE(root) -
1765                      btrfs_item_offset_nr(right, push_items -1);
1766
1767         copy_extent_buffer(left, right, btrfs_leaf_data(left) +
1768                      leaf_data_end(root, left) - push_space,
1769                      btrfs_leaf_data(right) +
1770                      btrfs_item_offset_nr(right, push_items - 1),
1771                      push_space);
1772         old_left_nritems = btrfs_header_nritems(left);
1773         BUG_ON(old_left_nritems < 0);
1774
1775         old_left_item_size = btrfs_item_offset_nr(left, old_left_nritems - 1);
1776         for (i = old_left_nritems; i < old_left_nritems + push_items; i++) {
1777                 u32 ioff;
1778
1779                 item = btrfs_item_nr(left, i);
1780                 if (!left->map_token) {
1781                         map_extent_buffer(left, (unsigned long)item,
1782                                         sizeof(struct btrfs_item),
1783                                         &left->map_token, &left->kaddr,
1784                                         &left->map_start, &left->map_len,
1785                                         KM_USER1);
1786                 }
1787
1788                 ioff = btrfs_item_offset(left, item);
1789                 btrfs_set_item_offset(left, item,
1790                       ioff - (BTRFS_LEAF_DATA_SIZE(root) - old_left_item_size));
1791         }
1792         btrfs_set_header_nritems(left, old_left_nritems + push_items);
1793         if (left->map_token) {
1794                 unmap_extent_buffer(left, left->map_token, KM_USER1);
1795                 left->map_token = NULL;
1796         }
1797
1798         /* fixup right node */
1799         if (push_items > right_nritems) {
1800                 printk("push items %d nr %u\n", push_items, right_nritems);
1801                 WARN_ON(1);
1802         }
1803
1804         if (push_items < right_nritems) {
1805                 push_space = btrfs_item_offset_nr(right, push_items - 1) -
1806                                                   leaf_data_end(root, right);
1807                 memmove_extent_buffer(right, btrfs_leaf_data(right) +
1808                                       BTRFS_LEAF_DATA_SIZE(root) - push_space,
1809                                       btrfs_leaf_data(right) +
1810                                       leaf_data_end(root, right), push_space);
1811
1812                 memmove_extent_buffer(right, btrfs_item_nr_offset(0),
1813                               btrfs_item_nr_offset(push_items),
1814                              (btrfs_header_nritems(right) - push_items) *
1815                              sizeof(struct btrfs_item));
1816         }
1817         right_nritems -= push_items;
1818         btrfs_set_header_nritems(right, right_nritems);
1819         push_space = BTRFS_LEAF_DATA_SIZE(root);
1820         for (i = 0; i < right_nritems; i++) {
1821                 item = btrfs_item_nr(right, i);
1822
1823                 if (!right->map_token) {
1824                         map_extent_buffer(right, (unsigned long)item,
1825                                         sizeof(struct btrfs_item),
1826                                         &right->map_token, &right->kaddr,
1827                                         &right->map_start, &right->map_len,
1828                                         KM_USER1);
1829                 }
1830
1831                 push_space = push_space - btrfs_item_size(right, item);
1832                 btrfs_set_item_offset(right, item, push_space);
1833         }
1834         if (right->map_token) {
1835                 unmap_extent_buffer(right, right->map_token, KM_USER1);
1836                 right->map_token = NULL;
1837         }
1838
1839         btrfs_mark_buffer_dirty(left);
1840         if (right_nritems)
1841                 btrfs_mark_buffer_dirty(right);
1842
1843         btrfs_item_key(right, &disk_key, 0);
1844         wret = fixup_low_keys(trans, root, path, &disk_key, 1);
1845         if (wret)
1846                 ret = wret;
1847
1848         /* then fixup the leaf pointer in the path */
1849         if (path->slots[0] < push_items) {
1850                 path->slots[0] += old_left_nritems;
1851                 free_extent_buffer(path->nodes[0]);
1852                 path->nodes[0] = left;
1853                 path->slots[1] -= 1;
1854         } else {
1855                 free_extent_buffer(left);
1856                 path->slots[0] -= push_items;
1857         }
1858         BUG_ON(path->slots[0] < 0);
1859         return ret;
1860 }
1861
1862 /*
1863  * split the path's leaf in two, making sure there is at least data_size
1864  * available for the resulting leaf level of the path.
1865  *
1866  * returns 0 if all went well and < 0 on failure.
1867  */
1868 static int split_leaf(struct btrfs_trans_handle *trans, struct btrfs_root
1869                       *root, struct btrfs_key *ins_key,
1870                       struct btrfs_path *path, int data_size, int extend)
1871 {
1872         u64 root_gen;
1873         struct extent_buffer *l;
1874         u32 nritems;
1875         int mid;
1876         int slot;
1877         struct extent_buffer *right;
1878         int space_needed = data_size + sizeof(struct btrfs_item);
1879         int data_copy_size;
1880         int rt_data_off;
1881         int i;
1882         int ret = 0;
1883         int wret;
1884         int double_split;
1885         int num_doubles = 0;
1886         struct btrfs_disk_key disk_key;
1887
1888         if (extend)
1889                 space_needed = data_size;
1890
1891         if (root->ref_cows)
1892                 root_gen = trans->transid;
1893         else
1894                 root_gen = 0;
1895
1896         /* first try to make some room by pushing left and right */
1897         if (ins_key->type != BTRFS_DIR_ITEM_KEY) {
1898                 wret = push_leaf_right(trans, root, path, data_size, 0);
1899                 if (wret < 0) {
1900                         return wret;
1901                 }
1902                 if (wret) {
1903                         wret = push_leaf_left(trans, root, path, data_size, 0);
1904                         if (wret < 0)
1905                                 return wret;
1906                 }
1907                 l = path->nodes[0];
1908
1909                 /* did the pushes work? */
1910                 if (btrfs_leaf_free_space(root, l) >= space_needed)
1911                         return 0;
1912         }
1913
1914         if (!path->nodes[1]) {
1915                 ret = insert_new_root(trans, root, path, 1);
1916                 if (ret)
1917                         return ret;
1918         }
1919 again:
1920         double_split = 0;
1921         l = path->nodes[0];
1922         slot = path->slots[0];
1923         nritems = btrfs_header_nritems(l);
1924         mid = (nritems + 1)/ 2;
1925
1926         btrfs_item_key(l, &disk_key, 0);
1927
1928         right = __btrfs_alloc_free_block(trans, root, root->leafsize,
1929                                          root->root_key.objectid,
1930                                          root_gen, disk_key.objectid, 0,
1931                                          l->start, 0);
1932         if (IS_ERR(right))
1933                 return PTR_ERR(right);
1934
1935         memset_extent_buffer(right, 0, 0, sizeof(struct btrfs_header));
1936         btrfs_set_header_bytenr(right, right->start);
1937         btrfs_set_header_generation(right, trans->transid);
1938         btrfs_set_header_owner(right, root->root_key.objectid);
1939         btrfs_set_header_level(right, 0);
1940         write_extent_buffer(right, root->fs_info->fsid,
1941                             (unsigned long)btrfs_header_fsid(right),
1942                             BTRFS_FSID_SIZE);
1943         if (mid <= slot) {
1944                 if (nritems == 1 ||
1945                     leaf_space_used(l, mid, nritems - mid) + space_needed >
1946                         BTRFS_LEAF_DATA_SIZE(root)) {
1947                         if (slot >= nritems) {
1948                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1949                                 btrfs_set_header_nritems(right, 0);
1950                                 wret = insert_ptr(trans, root, path,
1951                                                   &disk_key, right->start,
1952                                                   path->slots[1] + 1, 1);
1953                                 if (wret)
1954                                         ret = wret;
1955                                 free_extent_buffer(path->nodes[0]);
1956                                 path->nodes[0] = right;
1957                                 path->slots[0] = 0;
1958                                 path->slots[1] += 1;
1959                                 return ret;
1960                         }
1961                         mid = slot;
1962                         if (mid != nritems &&
1963                             leaf_space_used(l, mid, nritems - mid) +
1964                             space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1965                                 double_split = 1;
1966                         }
1967                 }
1968         } else {
1969                 if (leaf_space_used(l, 0, mid + 1) + space_needed >
1970                         BTRFS_LEAF_DATA_SIZE(root)) {
1971                         if (!extend && slot == 0) {
1972                                 btrfs_cpu_key_to_disk(&disk_key, ins_key);
1973                                 btrfs_set_header_nritems(right, 0);
1974                                 wret = insert_ptr(trans, root, path,
1975                                                   &disk_key,
1976                                                   right->start,
1977                                                   path->slots[1], 1);
1978                                 if (wret)
1979                                         ret = wret;
1980                                 free_extent_buffer(path->nodes[0]);
1981                                 path->nodes[0] = right;
1982                                 path->slots[0] = 0;
1983                                 if (path->slots[1] == 0) {
1984                                         wret = fixup_low_keys(trans, root,
1985                                                    path, &disk_key, 1);
1986                                         if (wret)
1987                                                 ret = wret;
1988                                 }
1989                                 return ret;
1990                         } else if (extend && slot == 0) {
1991                                 mid = 1;
1992                         } else {
1993                                 mid = slot;
1994                                 if (mid != nritems &&
1995                                     leaf_space_used(l, mid, nritems - mid) +
1996                                     space_needed > BTRFS_LEAF_DATA_SIZE(root)) {
1997                                         double_split = 1;
1998                                 }
1999                         }
2000                 }
2001         }
2002         nritems = nritems - mid;
2003         btrfs_set_header_nritems(right, nritems);
2004         data_copy_size = btrfs_item_end_nr(l, mid) - leaf_data_end(root, l);
2005
2006         copy_extent_buffer(right, l, btrfs_item_nr_offset(0),
2007                            btrfs_item_nr_offset(mid),
2008                            nritems * sizeof(struct btrfs_item));
2009
2010         copy_extent_buffer(right, l,
2011                      btrfs_leaf_data(right) + BTRFS_LEAF_DATA_SIZE(root) -
2012                      data_copy_size, btrfs_leaf_data(l) +
2013                      leaf_data_end(root, l), data_copy_size);
2014
2015         rt_data_off = BTRFS_LEAF_DATA_SIZE(root) -
2016                       btrfs_item_end_nr(l, mid);
2017
2018         for (i = 0; i < nritems; i++) {
2019                 struct btrfs_item *item = btrfs_item_nr(right, i);
2020                 u32 ioff;
2021
2022                 if (!right->map_token) {
2023                         map_extent_buffer(right, (unsigned long)item,
2024                                         sizeof(struct btrfs_item),
2025                                         &right->map_token, &right->kaddr,
2026                                         &right->map_start, &right->map_len,
2027                                         KM_USER1);
2028                 }
2029
2030                 ioff = btrfs_item_offset(right, item);
2031                 btrfs_set_item_offset(right, item, ioff + rt_data_off);
2032         }
2033
2034         if (right->map_token) {
2035                 unmap_extent_buffer(right, right->map_token, KM_USER1);
2036                 right->map_token = NULL;
2037         }
2038
2039         btrfs_set_header_nritems(l, mid);
2040         ret = 0;
2041         btrfs_item_key(right, &disk_key, 0);
2042         wret = insert_ptr(trans, root, path, &disk_key, right->start,
2043                           path->slots[1] + 1, 1);
2044         if (wret)
2045                 ret = wret;
2046
2047         btrfs_mark_buffer_dirty(right);
2048         btrfs_mark_buffer_dirty(l);
2049         BUG_ON(path->slots[0] != slot);
2050
2051         if (mid <= slot) {
2052                 free_extent_buffer(path->nodes[0]);
2053                 path->nodes[0] = right;
2054                 path->slots[0] -= mid;
2055                 path->slots[1] += 1;
2056         } else
2057                 free_extent_buffer(right);
2058
2059         BUG_ON(path->slots[0] < 0);
2060
2061         if (double_split) {
2062                 BUG_ON(num_doubles != 0);
2063                 num_doubles++;
2064                 goto again;
2065         }
2066         return ret;
2067 }
2068
2069 int btrfs_truncate_item(struct btrfs_trans_handle *trans,
2070                         struct btrfs_root *root,
2071                         struct btrfs_path *path,
2072                         u32 new_size, int from_end)
2073 {
2074         int ret = 0;
2075         int slot;
2076         int slot_orig;
2077         struct extent_buffer *leaf;
2078         struct btrfs_item *item;
2079         u32 nritems;
2080         unsigned int data_end;
2081         unsigned int old_data_start;
2082         unsigned int old_size;
2083         unsigned int size_diff;
2084         int i;
2085
2086         slot_orig = path->slots[0];
2087         leaf = path->nodes[0];
2088         slot = path->slots[0];
2089
2090         old_size = btrfs_item_size_nr(leaf, slot);
2091         if (old_size == new_size)
2092                 return 0;
2093
2094         nritems = btrfs_header_nritems(leaf);
2095         data_end = leaf_data_end(root, leaf);
2096
2097         old_data_start = btrfs_item_offset_nr(leaf, slot);
2098
2099         size_diff = old_size - new_size;
2100
2101         BUG_ON(slot < 0);
2102         BUG_ON(slot >= nritems);
2103
2104         /*
2105          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2106          */
2107         /* first correct the data pointers */
2108         for (i = slot; i < nritems; i++) {
2109                 u32 ioff;
2110                 item = btrfs_item_nr(leaf, i);
2111
2112                 if (!leaf->map_token) {
2113                         map_extent_buffer(leaf, (unsigned long)item,
2114                                         sizeof(struct btrfs_item),
2115                                         &leaf->map_token, &leaf->kaddr,
2116                                         &leaf->map_start, &leaf->map_len,
2117                                         KM_USER1);
2118                 }
2119
2120                 ioff = btrfs_item_offset(leaf, item);
2121                 btrfs_set_item_offset(leaf, item, ioff + size_diff);
2122         }
2123
2124         if (leaf->map_token) {
2125                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2126                 leaf->map_token = NULL;
2127         }
2128
2129         /* shift the data */
2130         if (from_end) {
2131                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2132                               data_end + size_diff, btrfs_leaf_data(leaf) +
2133                               data_end, old_data_start + new_size - data_end);
2134         } else {
2135                 struct btrfs_disk_key disk_key;
2136                 u64 offset;
2137
2138                 btrfs_item_key(leaf, &disk_key, slot);
2139
2140                 if (btrfs_disk_key_type(&disk_key) == BTRFS_EXTENT_DATA_KEY) {
2141                         unsigned long ptr;
2142                         struct btrfs_file_extent_item *fi;
2143
2144                         fi = btrfs_item_ptr(leaf, slot,
2145                                             struct btrfs_file_extent_item);
2146                         fi = (struct btrfs_file_extent_item *)(
2147                              (unsigned long)fi - size_diff);
2148
2149                         if (btrfs_file_extent_type(leaf, fi) ==
2150                             BTRFS_FILE_EXTENT_INLINE) {
2151                                 ptr = btrfs_item_ptr_offset(leaf, slot);
2152                                 memmove_extent_buffer(leaf, ptr,
2153                                         (unsigned long)fi,
2154                                         offsetof(struct btrfs_file_extent_item,
2155                                                  disk_bytenr));
2156                         }
2157                 }
2158
2159                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2160                               data_end + size_diff, btrfs_leaf_data(leaf) +
2161                               data_end, old_data_start - data_end);
2162
2163                 offset = btrfs_disk_key_offset(&disk_key);
2164                 btrfs_set_disk_key_offset(&disk_key, offset + size_diff);
2165                 btrfs_set_item_key(leaf, &disk_key, slot);
2166                 if (slot == 0)
2167                         fixup_low_keys(trans, root, path, &disk_key, 1);
2168         }
2169
2170         item = btrfs_item_nr(leaf, slot);
2171         btrfs_set_item_size(leaf, item, new_size);
2172         btrfs_mark_buffer_dirty(leaf);
2173
2174         ret = 0;
2175         if (btrfs_leaf_free_space(root, leaf) < 0) {
2176                 btrfs_print_leaf(root, leaf);
2177                 BUG();
2178         }
2179         return ret;
2180 }
2181
2182 int btrfs_extend_item(struct btrfs_trans_handle *trans,
2183                       struct btrfs_root *root, struct btrfs_path *path,
2184                       u32 data_size)
2185 {
2186         int ret = 0;
2187         int slot;
2188         int slot_orig;
2189         struct extent_buffer *leaf;
2190         struct btrfs_item *item;
2191         u32 nritems;
2192         unsigned int data_end;
2193         unsigned int old_data;
2194         unsigned int old_size;
2195         int i;
2196
2197         slot_orig = path->slots[0];
2198         leaf = path->nodes[0];
2199
2200         nritems = btrfs_header_nritems(leaf);
2201         data_end = leaf_data_end(root, leaf);
2202
2203         if (btrfs_leaf_free_space(root, leaf) < data_size) {
2204                 btrfs_print_leaf(root, leaf);
2205                 BUG();
2206         }
2207         slot = path->slots[0];
2208         old_data = btrfs_item_end_nr(leaf, slot);
2209
2210         BUG_ON(slot < 0);
2211         if (slot >= nritems) {
2212                 btrfs_print_leaf(root, leaf);
2213                 printk("slot %d too large, nritems %d\n", slot, nritems);
2214                 BUG_ON(1);
2215         }
2216
2217         /*
2218          * item0..itemN ... dataN.offset..dataN.size .. data0.size
2219          */
2220         /* first correct the data pointers */
2221         for (i = slot; i < nritems; i++) {
2222                 u32 ioff;
2223                 item = btrfs_item_nr(leaf, i);
2224
2225                 if (!leaf->map_token) {
2226                         map_extent_buffer(leaf, (unsigned long)item,
2227                                         sizeof(struct btrfs_item),
2228                                         &leaf->map_token, &leaf->kaddr,
2229                                         &leaf->map_start, &leaf->map_len,
2230                                         KM_USER1);
2231                 }
2232                 ioff = btrfs_item_offset(leaf, item);
2233                 btrfs_set_item_offset(leaf, item, ioff - data_size);
2234         }
2235
2236         if (leaf->map_token) {
2237                 unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2238                 leaf->map_token = NULL;
2239         }
2240
2241         /* shift the data */
2242         memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2243                       data_end - data_size, btrfs_leaf_data(leaf) +
2244                       data_end, old_data - data_end);
2245
2246         data_end = old_data;
2247         old_size = btrfs_item_size_nr(leaf, slot);
2248         item = btrfs_item_nr(leaf, slot);
2249         btrfs_set_item_size(leaf, item, old_size + data_size);
2250         btrfs_mark_buffer_dirty(leaf);
2251
2252         ret = 0;
2253         if (btrfs_leaf_free_space(root, leaf) < 0) {
2254                 btrfs_print_leaf(root, leaf);
2255                 BUG();
2256         }
2257         return ret;
2258 }
2259
2260 /*
2261  * Given a key and some data, insert an item into the tree.
2262  * This does all the path init required, making room in the tree if needed.
2263  */
2264 int btrfs_insert_empty_item(struct btrfs_trans_handle *trans,
2265                             struct btrfs_root *root,
2266                             struct btrfs_path *path,
2267                             struct btrfs_key *cpu_key, u32 data_size)
2268 {
2269         struct extent_buffer *leaf;
2270         struct btrfs_item *item;
2271         int ret = 0;
2272         int slot;
2273         int slot_orig;
2274         u32 nritems;
2275         unsigned int data_end;
2276         struct btrfs_disk_key disk_key;
2277
2278         btrfs_cpu_key_to_disk(&disk_key, cpu_key);
2279
2280         /* create a root if there isn't one */
2281         if (!root->node)
2282                 BUG();
2283
2284         ret = btrfs_search_slot(trans, root, cpu_key, path, data_size, 1);
2285         if (ret == 0) {
2286                 return -EEXIST;
2287         }
2288         if (ret < 0)
2289                 goto out;
2290
2291         slot_orig = path->slots[0];
2292         leaf = path->nodes[0];
2293
2294         nritems = btrfs_header_nritems(leaf);
2295         data_end = leaf_data_end(root, leaf);
2296
2297         if (btrfs_leaf_free_space(root, leaf) <
2298             sizeof(struct btrfs_item) + data_size) {
2299                 btrfs_print_leaf(root, leaf);
2300                 printk("not enough freespace need %u have %d\n",
2301                        data_size, btrfs_leaf_free_space(root, leaf));
2302                 BUG();
2303         }
2304
2305         slot = path->slots[0];
2306         BUG_ON(slot < 0);
2307
2308         if (slot != nritems) {
2309                 int i;
2310                 unsigned int old_data = btrfs_item_end_nr(leaf, slot);
2311
2312                 if (old_data < data_end) {
2313                         btrfs_print_leaf(root, leaf);
2314                         printk("slot %d old_data %d data_end %d\n",
2315                                slot, old_data, data_end);
2316                         BUG_ON(1);
2317                 }
2318                 /*
2319                  * item0..itemN ... dataN.offset..dataN.size .. data0.size
2320                  */
2321                 /* first correct the data pointers */
2322                 WARN_ON(leaf->map_token);
2323                 for (i = slot; i < nritems; i++) {
2324                         u32 ioff;
2325
2326                         item = btrfs_item_nr(leaf, i);
2327                         if (!leaf->map_token) {
2328                                 map_extent_buffer(leaf, (unsigned long)item,
2329                                         sizeof(struct btrfs_item),
2330                                         &leaf->map_token, &leaf->kaddr,
2331                                         &leaf->map_start, &leaf->map_len,
2332                                         KM_USER1);
2333                         }
2334
2335                         ioff = btrfs_item_offset(leaf, item);
2336                         btrfs_set_item_offset(leaf, item, ioff - data_size);
2337                 }
2338                 if (leaf->map_token) {
2339                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2340                         leaf->map_token = NULL;
2341                 }
2342
2343                 /* shift the items */
2344                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot + 1),
2345                               btrfs_item_nr_offset(slot),
2346                               (nritems - slot) * sizeof(struct btrfs_item));
2347
2348                 /* shift the data */
2349                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2350                               data_end - data_size, btrfs_leaf_data(leaf) +
2351                               data_end, old_data - data_end);
2352                 data_end = old_data;
2353         }
2354
2355         /* setup the item for the new data */
2356         btrfs_set_item_key(leaf, &disk_key, slot);
2357         item = btrfs_item_nr(leaf, slot);
2358         btrfs_set_item_offset(leaf, item, data_end - data_size);
2359         btrfs_set_item_size(leaf, item, data_size);
2360         btrfs_set_header_nritems(leaf, nritems + 1);
2361         btrfs_mark_buffer_dirty(leaf);
2362
2363         ret = 0;
2364         if (slot == 0)
2365                 ret = fixup_low_keys(trans, root, path, &disk_key, 1);
2366
2367         if (btrfs_leaf_free_space(root, leaf) < 0) {
2368                 btrfs_print_leaf(root, leaf);
2369                 BUG();
2370         }
2371 out:
2372         return ret;
2373 }
2374
2375 /*
2376  * Given a key and some data, insert an item into the tree.
2377  * This does all the path init required, making room in the tree if needed.
2378  */
2379 int btrfs_insert_item(struct btrfs_trans_handle *trans, struct btrfs_root
2380                       *root, struct btrfs_key *cpu_key, void *data, u32
2381                       data_size)
2382 {
2383         int ret = 0;
2384         struct btrfs_path *path;
2385         struct extent_buffer *leaf;
2386         unsigned long ptr;
2387
2388         path = btrfs_alloc_path();
2389         BUG_ON(!path);
2390         ret = btrfs_insert_empty_item(trans, root, path, cpu_key, data_size);
2391         if (!ret) {
2392                 leaf = path->nodes[0];
2393                 ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
2394                 write_extent_buffer(leaf, data, ptr, data_size);
2395                 btrfs_mark_buffer_dirty(leaf);
2396         }
2397         btrfs_free_path(path);
2398         return ret;
2399 }
2400
2401 /*
2402  * delete the pointer from a given node.
2403  *
2404  * If the delete empties a node, the node is removed from the tree,
2405  * continuing all the way the root if required.  The root is converted into
2406  * a leaf if all the nodes are emptied.
2407  */
2408 static int del_ptr(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2409                    struct btrfs_path *path, int level, int slot)
2410 {
2411         struct extent_buffer *parent = path->nodes[level];
2412         u32 nritems;
2413         int ret = 0;
2414         int wret;
2415
2416         nritems = btrfs_header_nritems(parent);
2417         if (slot != nritems -1) {
2418                 memmove_extent_buffer(parent,
2419                               btrfs_node_key_ptr_offset(slot),
2420                               btrfs_node_key_ptr_offset(slot + 1),
2421                               sizeof(struct btrfs_key_ptr) *
2422                               (nritems - slot - 1));
2423         }
2424         nritems--;
2425         btrfs_set_header_nritems(parent, nritems);
2426         if (nritems == 0 && parent == root->node) {
2427                 BUG_ON(btrfs_header_level(root->node) != 1);
2428                 /* just turn the root into a leaf and break */
2429                 btrfs_set_header_level(root->node, 0);
2430         } else if (slot == 0) {
2431                 struct btrfs_disk_key disk_key;
2432
2433                 btrfs_node_key(parent, &disk_key, 0);
2434                 wret = fixup_low_keys(trans, root, path, &disk_key, level + 1);
2435                 if (wret)
2436                         ret = wret;
2437         }
2438         btrfs_mark_buffer_dirty(parent);
2439         return ret;
2440 }
2441
2442 /*
2443  * delete the item at the leaf level in path.  If that empties
2444  * the leaf, remove it from the tree
2445  */
2446 int btrfs_del_item(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2447                    struct btrfs_path *path)
2448 {
2449         int slot;
2450         struct extent_buffer *leaf;
2451         struct btrfs_item *item;
2452         int doff;
2453         int dsize;
2454         int ret = 0;
2455         int wret;
2456         u32 nritems;
2457
2458         leaf = path->nodes[0];
2459         slot = path->slots[0];
2460         doff = btrfs_item_offset_nr(leaf, slot);
2461         dsize = btrfs_item_size_nr(leaf, slot);
2462         nritems = btrfs_header_nritems(leaf);
2463
2464         if (slot != nritems - 1) {
2465                 int i;
2466                 int data_end = leaf_data_end(root, leaf);
2467
2468                 memmove_extent_buffer(leaf, btrfs_leaf_data(leaf) +
2469                               data_end + dsize,
2470                               btrfs_leaf_data(leaf) + data_end,
2471                               doff - data_end);
2472
2473                 for (i = slot + 1; i < nritems; i++) {
2474                         u32 ioff;
2475
2476                         item = btrfs_item_nr(leaf, i);
2477                         if (!leaf->map_token) {
2478                                 map_extent_buffer(leaf, (unsigned long)item,
2479                                         sizeof(struct btrfs_item),
2480                                         &leaf->map_token, &leaf->kaddr,
2481                                         &leaf->map_start, &leaf->map_len,
2482                                         KM_USER1);
2483                         }
2484                         ioff = btrfs_item_offset(leaf, item);
2485                         btrfs_set_item_offset(leaf, item, ioff + dsize);
2486                 }
2487
2488                 if (leaf->map_token) {
2489                         unmap_extent_buffer(leaf, leaf->map_token, KM_USER1);
2490                         leaf->map_token = NULL;
2491                 }
2492
2493                 memmove_extent_buffer(leaf, btrfs_item_nr_offset(slot),
2494                               btrfs_item_nr_offset(slot + 1),
2495                               sizeof(struct btrfs_item) *
2496                               (nritems - slot - 1));
2497         }
2498         btrfs_set_header_nritems(leaf, nritems - 1);
2499         nritems--;
2500
2501         /* delete the leaf if we've emptied it */
2502         if (nritems == 0) {
2503                 if (leaf == root->node) {
2504                         btrfs_set_header_level(leaf, 0);
2505                 } else {
2506                         u64 root_gen = btrfs_header_generation(path->nodes[1]);
2507                         clean_tree_block(trans, root, leaf);
2508                         wait_on_tree_block_writeback(root, leaf);
2509                         wret = del_ptr(trans, root, path, 1, path->slots[1]);
2510                         if (wret)
2511                                 ret = wret;
2512                         wret = btrfs_free_extent(trans, root,
2513                                          leaf->start, leaf->len,
2514                                          btrfs_header_owner(path->nodes[1]),
2515                                          root_gen, 0, 0, 1);
2516                         if (wret)
2517                                 ret = wret;
2518                 }
2519         } else {
2520                 int used = leaf_space_used(leaf, 0, nritems);
2521                 if (slot == 0) {
2522                         struct btrfs_disk_key disk_key;
2523
2524                         btrfs_item_key(leaf, &disk_key, 0);
2525                         wret = fixup_low_keys(trans, root, path,
2526                                               &disk_key, 1);
2527                         if (wret)
2528                                 ret = wret;
2529                 }
2530
2531                 /* delete the leaf if it is mostly empty */
2532                 if (used < BTRFS_LEAF_DATA_SIZE(root) / 3) {
2533                         /* push_leaf_left fixes the path.
2534                          * make sure the path still points to our leaf
2535                          * for possible call to del_ptr below
2536                          */
2537                         slot = path->slots[1];
2538                         extent_buffer_get(leaf);
2539
2540                         wret = push_leaf_right(trans, root, path, 1, 1);
2541                         if (wret < 0 && wret != -ENOSPC)
2542                                 ret = wret;
2543
2544                         if (path->nodes[0] == leaf &&
2545                             btrfs_header_nritems(leaf)) {
2546                                 wret = push_leaf_left(trans, root, path, 1, 1);
2547                                 if (wret < 0 && wret != -ENOSPC)
2548                                         ret = wret;
2549                         }
2550
2551                         if (btrfs_header_nritems(leaf) == 0) {
2552                                 u64 root_gen;
2553                                 u64 bytenr = leaf->start;
2554                                 u32 blocksize = leaf->len;
2555
2556                                 root_gen = btrfs_header_generation(
2557                                                            path->nodes[1]);
2558
2559                                 clean_tree_block(trans, root, leaf);
2560                                 wait_on_tree_block_writeback(root, leaf);
2561
2562                                 wret = del_ptr(trans, root, path, 1, slot);
2563                                 if (wret)
2564                                         ret = wret;
2565
2566                                 free_extent_buffer(leaf);
2567                                 wret = btrfs_free_extent(trans, root, bytenr,
2568                                              blocksize,
2569                                              btrfs_header_owner(path->nodes[1]),
2570                                              root_gen, 0, 0, 1);
2571                                 if (wret)
2572                                         ret = wret;
2573                         } else {
2574                                 btrfs_mark_buffer_dirty(leaf);
2575                                 free_extent_buffer(leaf);
2576                         }
2577                 } else {
2578                         btrfs_mark_buffer_dirty(leaf);
2579                 }
2580         }
2581         return ret;
2582 }
2583
2584 /*
2585  * walk up the tree as far as required to find the previous leaf.
2586  * returns 0 if it found something or 1 if there are no lesser leaves.
2587  * returns < 0 on io errors.
2588  */
2589 int btrfs_prev_leaf(struct btrfs_root *root, struct btrfs_path *path)
2590 {
2591         int slot;
2592         int level = 1;
2593         u64 bytenr;
2594         struct extent_buffer *c;
2595         struct extent_buffer *next = NULL;
2596
2597         while(level < BTRFS_MAX_LEVEL) {
2598                 if (!path->nodes[level])
2599                         return 1;
2600
2601                 slot = path->slots[level];
2602                 c = path->nodes[level];
2603                 if (slot == 0) {
2604                         level++;
2605                         if (level == BTRFS_MAX_LEVEL)
2606                                 return 1;
2607                         continue;
2608                 }
2609                 slot--;
2610
2611                 bytenr = btrfs_node_blockptr(c, slot);
2612                 if (next)
2613                         free_extent_buffer(next);
2614
2615                 if (path->reada < 0)
2616                         reada_for_search(root, path, level, slot);
2617
2618                 next = read_tree_block(root, bytenr,
2619                                        btrfs_level_size(root, level - 1));
2620                 break;
2621         }
2622         path->slots[level] = slot;
2623         while(1) {
2624                 level--;
2625                 c = path->nodes[level];
2626                 free_extent_buffer(c);
2627                 path->nodes[level] = next;
2628                 path->slots[level] = 0;
2629                 if (!level)
2630                         break;
2631                 if (path->reada)
2632                         reada_for_search(root, path, level, 0);
2633                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2634                                        btrfs_level_size(root, level - 1));
2635         }
2636         return 0;
2637 }
2638
2639 /*
2640  * walk up the tree as far as required to find the next leaf.
2641  * returns 0 if it found something or 1 if there are no greater leaves.
2642  * returns < 0 on io errors.
2643  */
2644 int btrfs_next_leaf(struct btrfs_root *root, struct btrfs_path *path)
2645 {
2646         int slot;
2647         int level = 1;
2648         u64 bytenr;
2649         struct extent_buffer *c;
2650         struct extent_buffer *next = NULL;
2651
2652         while(level < BTRFS_MAX_LEVEL) {
2653                 if (!path->nodes[level])
2654                         return 1;
2655
2656                 slot = path->slots[level] + 1;
2657                 c = path->nodes[level];
2658                 if (slot >= btrfs_header_nritems(c)) {
2659                         level++;
2660                         if (level == BTRFS_MAX_LEVEL)
2661                                 return 1;
2662                         continue;
2663                 }
2664
2665                 bytenr = btrfs_node_blockptr(c, slot);
2666                 if (next)
2667                         free_extent_buffer(next);
2668
2669                 if (path->reada)
2670                         reada_for_search(root, path, level, slot);
2671
2672                 next = read_tree_block(root, bytenr,
2673                                        btrfs_level_size(root, level -1));
2674                 break;
2675         }
2676         path->slots[level] = slot;
2677         while(1) {
2678                 level--;
2679                 c = path->nodes[level];
2680                 free_extent_buffer(c);
2681                 path->nodes[level] = next;
2682                 path->slots[level] = 0;
2683                 if (!level)
2684                         break;
2685                 if (path->reada)
2686                         reada_for_search(root, path, level, 0);
2687                 next = read_tree_block(root, btrfs_node_blockptr(next, 0),
2688                                        btrfs_level_size(root, level - 1));
2689         }
2690         return 0;
2691 }